Deep Learning-Based Lung Segmentation for Multi-Modal Imaging Data using Attention Residual U-Net

Precise identification of lung regions in CT scans is essential for lung cancer diagnosis, staging, and quantitative assessment. Inaccurate or inconsistent delineation can compromise measurements and affect clinical decisions. Traditional segmentation methods, including standard U-Net architectures, often struggle when confronted with variations in imaging protocols or abnormal lung appearances caused by disease. To address these limitations, this study proposes a deep learning–based framework using the Attention Residual U-Net (ARU-Net) for generating accurate lung masks across diverse DICOM datasets. ARU-Net strengthens feature propagation through residual connections while its attention mechanism enables the network to focus more effectively on relevant lung structures and suppress background interference. The model is initially trained on the Kaggle lung segmentation datasets (LUNA16 and DSB2017), which provide expert -annotated 2D CT slices, and later applied to multi-institutional DICOM scans from The Cancer Imaging Archive (TCIA), including CT and PET -CT studies. Pre-processing steps such as intensity normalization and histogram matching are incorporated to enhance domain consistency. The resulting lung masks are produced as 3D volumes and DICOM SEG overlays to support further clinical tasks, including lesion extraction, TNM staging, and percentile density analysis. Experimental results show that the proposed method outperforms conventional U-Net models in segmentation accuracy, robustness, and downstream clinical applicability.